OpenAI and Altera: Creating Collaborative Digital Humans with GPT-4o
Altera has developed "digital humans"—autonomous AI agents capable of long-term collaboration and emotional interaction with people—using a brain-inspired architecture powered by OpenAI's GPT-4o. This advancement allows agents to maintain stability and cognitive function over extended periods, overcoming the common industry challenge of data degradation.
Overcoming Data Degradation in Autonomous Agents
AI agents that make real-time decisions over extended timeframes often suffer from data degradation, where the agent's own output becomes its future input, leading to a decline in data quality and performance. Altera identifies this is a critical barrier for agents intended to live autonomously for hours or longer.
To solve this, Altera integrated OpenAI's language models to maintain the integrity of decision-making processes. This integration allows agents to achieve longer and more complex interactions without the rapid performance decline typically seen in autonomous AI systems.
Brain-Inspired Architecture and Cognitive Simulation
Altera utilizes a parallel multi-module system designed to mimic the structure of the human brain, specifically the prefrontal cortex. This composite system runs various modules in parallel, each powered by OpenAI models, to simulate specific cognitive functions:
- Attention Bottleneck: Managing how the agent focuses on information.
- Working Memory: Maintaining short-term information for active tasks.
- Social Cognition: Enabling the agent to understand and interact with others.
According to CEO Dr. Robert Yang, this architecture allows agents to "think faster, process emotions, and ultimately build a long-term sense of self."
Capabilities and Current Benchmarks
As of mid-2024, Altera's digital humans can operate autonomously for up to four hours at a time, which the company states is a substantial increase over other AI models currently on the market.
Altera's first product demonstrates these capabilities through agents that can play and interact with humans in Minecraft, specifically utilizing a "Distinct Item Collection" benchmark to measure success in virtual environments. This serves as a foundation for future multi-agent simulations, including the creation of entire digital societies to test economic policies or advertisements.
Future Applications of Digital Humans
Altera envisions a future where AI agents transition from simple assistants to collaborative partners. The company identifies two primary trajectories for this technology:
- Productivity: Digital "coworkers" capable of collaborating for days or weeks to proactively solve problems.
- Simulation: Long-duration multi-agent worlds used to measure responses to economic policies and marketing strategies.